计算混淆矩阵的更快方法？

我正在计算图像语义分割的混淆矩阵，如下所示，这是一种非常冗长的方法：

def confusion_matrix(preds, labels, conf_m, sample_size):
    preds = normalize(preds,0.9) # returns [0,1] tensor
    preds = preds.flatten()
    labels = labels.flatten()
    for i in range(len(preds)):
        if preds[i]==1 and labels[i]==1:
            conf_m[0,0] += 1/(len(preds)*sample_size) # TP
        elif preds[i]==1 and labels[i]==0:
            conf_m[0,1] += 1/(len(preds)*sample_size) # FP
        elif preds[i]==0 and labels[i]==0:
            conf_m[1,0] += 1/(len(preds)*sample_size) # TN
        elif preds[i]==0 and labels[i]==1:
            conf_m[1,1] += 1/(len(preds)*sample_size) # FN 
    return conf_m

在预测循环中：

conf_m = torch.zeros(2,2) # two classes (object or no-object)
for img,label in enumerate(data):
    ...
    out = Net(img)
    conf_m = confusion_matrix(out, label, len(data))
    ...

是否有更快的方法（在 PyTorch 中）来有效计算图像语义分割输入样本的混淆矩阵？

我使用这两个函数来计算混淆矩阵（如其定义）sklearn):

# rewrite sklearn method to torch
def confusion_matrix_1(y_true, y_pred):
    N = max(max(y_true), max(y_pred)) + 1
    y_true = torch.tensor(y_true, dtype=torch.long)
    y_pred = torch.tensor(y_pred, dtype=torch.long)
    return torch.sparse.LongTensor(
        torch.stack([y_true, y_pred]), 
        torch.ones_like(y_true, dtype=torch.long),
        torch.Size([N, N])).to_dense()

# weird trick with bincount
def confusion_matrix_2(y_true, y_pred):
    N = max(max(y_true), max(y_pred)) + 1
    y_true = torch.tensor(y_true, dtype=torch.long)
    y_pred = torch.tensor(y_pred, dtype=torch.long)
    y = N * y_true + y_pred
    y = torch.bincount(y)
    if len(y) < N * N:
        y = torch.cat(y, torch.zeros(N * N - len(y), dtype=torch.long))
    y = y.reshape(N, N)
    return y

y_true = [2, 0, 2, 2, 0, 1]
y_pred = [0, 0, 2, 2, 0, 2]

confusion_matrix_1(y_true, y_pred)
# tensor([[2, 0, 0],
#         [0, 0, 1],
#         [1, 0, 2]])

在类数量较少的情况下，第二个函数速度更快。

%%timeit
confusion_matrix_1(y_true, y_pred)
# 102 µs ± 30.7 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

%%timeit
confusion_matrix_2(y_true, y_pred)
# 25 µs ± 149 ns per loop (mean ± std. dev. of 7 runs, 10000 loops each)